Abstract Extracting Animated Meshes with Adaptive Motion Estimation

Our main contribution is to show that for continuously evolving surface data the propagation by an adaptively estimated flow field can help to establish temporally coherent meshing of the dynamic surface.
1.1
Related work
A similar problem of creating a single coherent geometry representation for a number of shapes has been recently addressed by computer graphics researchers. For instance, when several range scans of the same individual are given, Allen et al.[ACP02] solve the problem of fitting a subdivision surface template to create a posable 3D model. That work was extended to handle the parameterization of whole-body data for multiple humans [ACP03]. In both cases, a sparse set of 3D markers is used for model registration. Related work on finding consistent parameterizations of dissimilar shapes by Praun et al. [PSS01] has also used user-defined feature markers in a multiresolution remeshing procedure. Neither of the studies mentioned above explicitly considers a continuously evolving surface. In this paper, we concentrate our effort on extracting a single animated mesh from a continuously evolving shape sequence. We do not assume the availability of markers or surface textures in the original data, with the hope that strong shape coherence will give enough information for extracting surface motion. We split our problem into the motion estimation and the mesh propagation parts. The motion estimation problem has long been studied in the computer vision community [BB81]. In particular, several efficient optical flow algorithms were introduced [SS96][Sim93]. We extend the multiscale optical flow algorithm of Simoncelli [Sim93] to work on adaptive signed distance volume representations and modify it to remove unneeded interactions between unrelated surface patches. Our motion estimation and surface fitting procedures require a volumetric shape representation, and an efficient conversion of input polygonal data is necessary for the overall good performance of the algorithm. We employ an adaptive shape representation similar to the ADF representation of Frisken et al. [FPRJ00]; due to some differences and to avoid confusion we generically call it ASDV (adaptive signed distance volume). We convert our input meshes into a sequence of ASDV datasets and use them for both motion estimation and surface fitting 666
ቤተ መጻሕፍቲ ባይዱ
Animated meshes are widely employed in character animation, visualization, and computational simulation applications. An animated mesh is a sequence of meshes with the same connectivity whose vertex positions change in time. This is a convenient representation for dynamically changing shapes, with many processing, rendering, and compression tasks easily handled. For instance, modern compression methods can take advantage of the temporal coherence present in animated mesh data, resulting in a very compact shape representation [IR03][KG]. Unfortunately, several recently introduced state of the art dynamic shape acquisition methods [SMP03][ZCS03] do not produce their result in the animated mesh form; rather a sequence of meshes of varying connectivity is produced. Volumetric morphing and isosurface extraction are also examples of applications that produce evolving surfaces that are not meshed consistently in time, and have changing connectivity and samVMV 2004
pling from frame to frame. Reconstruction of animated mesh sequences from such data is an important problem we aim to address in this paper.
1
Introduction
Abstract
We present an approach for extracting coherently sampled animated meshes from input sequences of incoherently sampled meshes representing a continuously evolving shape. Our approach is based on multiscale adaptive motion estimation procedure followed by propagation of a template mesh through time. An adaptive signed distance volumes are used as the principal shape representation, and a Bayesian optical flow algorithm is adapted to the surface setting with a modification that diminishes the interference between unrelated surface regions. Additionally, a parametric smoothing step is employed to improve the sampling coherence of the model. The result of the proposed procedure is a single animated mesh. We apply our approach to the human motion data.
Figure 1: Comparison of trajectories for three methods of particle propagation for the jump sequence: fitting, temporal prediction followed by fitting, our estimated flow followed by fitting. The left two propagations fail, the right one succeeds to preserve the surface sampling. We restrict our effort to the simpler scenario of unchanging topology both in the input data and in the desired output. A single mesh template of the same connectivity is propagated through time and is fitted to the input surface data in every frame of the sequence. The main challenge is to establish correspondence between consecutive shapes. Computing such correspondence relation constitutes the main contribution of this paper. We assume that the input surfaces are closed and without boundaries; the first step of the algorithm converts each input shape into an adaptive volumetric representation with a signed distance transform. Once a volumetric representation is obtained, we run a Bayesian motion estimation procedure similar to a differential optical flow approach from image processing. The resulting vector field defined on the surface is used for the initial propagation of the mesh template. The further fitting and parametric smoothing steps result in a temporally coherent animated mesh sequence. Stanford, USA, November 16–18, 2004